3 # ====================================================================
4 # Written by Andy Polyakov <appro@fy.chalmers.se> for the OpenSSL
5 # project. The module is, however, dual licensed under OpenSSL and
6 # CRYPTOGAMS licenses depending on where you obtain it. For further
7 # details see http://www.openssl.org/~appro/cryptogams/.
8 # ====================================================================
10 # Wrapper around 'rep montmul', VIA-specific instruction accessing
11 # PadLock Montgomery Multiplier. The wrapper is designed as drop-in
12 # replacement for OpenSSL bn_mul_mont [first implemented in 0.9.9].
14 # Below are interleaved outputs from 'openssl speed rsa dsa' for 4
15 # different software configurations on 1.5GHz VIA Esther processor.
16 # Lines marked with "software integer" denote performance of hand-
17 # coded integer-only assembler found in OpenSSL 0.9.7. "Software SSE2"
18 # refers to hand-coded SSE2 Montgomery multiplication procedure found
19 # OpenSSL 0.9.9. "Hardware VIA SDK" refers to padlock_pmm routine from
20 # Padlock SDK 2.0.1 available for download from VIA, which naturally
21 # utilizes the magic 'repz montmul' instruction. And finally "hardware
22 # this" refers to *this* implementation which also uses 'repz montmul'
24 # sign verify sign/s verify/s
25 # rsa 512 bits 0.001720s 0.000140s 581.4 7149.7 software integer
26 # rsa 512 bits 0.000690s 0.000086s 1450.3 11606.0 software SSE2
27 # rsa 512 bits 0.006136s 0.000201s 163.0 4974.5 hardware VIA SDK
28 # rsa 512 bits 0.000712s 0.000050s 1404.9 19858.5 hardware this
30 # rsa 1024 bits 0.008518s 0.000413s 117.4 2420.8 software integer
31 # rsa 1024 bits 0.004275s 0.000277s 233.9 3609.7 software SSE2
32 # rsa 1024 bits 0.012136s 0.000260s 82.4 3844.5 hardware VIA SDK
33 # rsa 1024 bits 0.002522s 0.000116s 396.5 8650.9 hardware this
35 # rsa 2048 bits 0.050101s 0.001371s 20.0 729.6 software integer
36 # rsa 2048 bits 0.030273s 0.001008s 33.0 991.9 software SSE2
37 # rsa 2048 bits 0.030833s 0.000976s 32.4 1025.1 hardware VIA SDK
38 # rsa 2048 bits 0.011879s 0.000342s 84.2 2921.7 hardware this
40 # rsa 4096 bits 0.327097s 0.004859s 3.1 205.8 software integer
41 # rsa 4096 bits 0.229318s 0.003859s 4.4 259.2 software SSE2
42 # rsa 4096 bits 0.233953s 0.003274s 4.3 305.4 hardware VIA SDK
43 # rsa 4096 bits 0.070493s 0.001166s 14.2 857.6 hardware this
45 # dsa 512 bits 0.001342s 0.001651s 745.2 605.7 software integer
46 # dsa 512 bits 0.000844s 0.000987s 1185.3 1013.1 software SSE2
47 # dsa 512 bits 0.001902s 0.002247s 525.6 444.9 hardware VIA SDK
48 # dsa 512 bits 0.000458s 0.000524s 2182.2 1909.1 hardware this
50 # dsa 1024 bits 0.003964s 0.004926s 252.3 203.0 software integer
51 # dsa 1024 bits 0.002686s 0.003166s 372.3 315.8 software SSE2
52 # dsa 1024 bits 0.002397s 0.002823s 417.1 354.3 hardware VIA SDK
53 # dsa 1024 bits 0.000978s 0.001170s 1022.2 855.0 hardware this
55 # dsa 2048 bits 0.013280s 0.016518s 75.3 60.5 software integer
56 # dsa 2048 bits 0.009911s 0.011522s 100.9 86.8 software SSE2
57 # dsa 2048 bits 0.009542s 0.011763s 104.8 85.0 hardware VIA SDK
58 # dsa 2048 bits 0.002884s 0.003352s 346.8 298.3 hardware this
60 # To give you some other reference point here is output for 2.4GHz P4
61 # running hand-coded SSE2 bn_mul_mont found in 0.9.9, i.e. "software
62 # SSE2" in above terms.
64 # rsa 512 bits 0.000407s 0.000047s 2454.2 21137.0
65 # rsa 1024 bits 0.002426s 0.000141s 412.1 7100.0
66 # rsa 2048 bits 0.015046s 0.000491s 66.5 2034.9
67 # rsa 4096 bits 0.109770s 0.002379s 9.1 420.3
68 # dsa 512 bits 0.000438s 0.000525s 2281.1 1904.1
69 # dsa 1024 bits 0.001346s 0.001595s 742.7 627.0
70 # dsa 2048 bits 0.004745s 0.005582s 210.7 179.1
73 # - VIA SDK leaves a *lot* of room for improvement (which this
74 # implementation successfully fills:-);
75 # - 'rep montmul' gives up to >3x performance improvement depending on
77 # - in terms of absolute performance it delivers approximately as much
78 # as modern out-of-order 32-bit cores [again, for longer keys].
80 push(@INC,".","../../perlasm");
83 &asm_init($ARGV[0],"via-mont.pl");
85 # int bn_mul_mont(BN_ULONG *rp, const BN_ULONG *ap, const BN_ULONG *bp, const BN_ULONG *np,const BN_ULONG *n0, int num);
86 $func="bn_mul_mont_padlock";
88 $pad=16*1; # amount of reserved bytes on top of every vector
91 $mZeroPrime=&DWP(0,"esp"); # these are specified by VIA
96 $scratch=&DWP(20,"esp");
97 $rp=&DWP(24,"esp"); # these are mine
99 # &DWP(32,"esp") # 32 byte scratch area
100 # &DWP(64+(4*$num+$pad)*0,"esp") # padded tp[num]
101 # &DWP(64+(4*$num+$pad)*1,"esp") # padded copy of ap[num]
102 # &DWP(64+(4*$num+$pad)*2,"esp") # padded copy of bp[num]
103 # &DWP(64+(4*$num+$pad)*2,"esp") # padded copy of np[num]
104 # Note that SDK suggests to unconditionally allocate 2K per vector. This
105 # has quite an impact on performance. It naturally depends on key length,
106 # but to give an example 1024 bit private RSA key operations suffer >30%
107 # penalty. I allocate only as much as actually required...
109 &function_begin($func);
111 &mov ("ecx",&wparam(5)); # num
112 # meet VIA's limitations for num [note that the specification
113 # expresses them in bits, while we work with amount of 32-bit words]
115 &jnz (&label("leave")); # num % 4 != 0
117 &jb (&label("leave")); # num < 8
119 &ja (&label("leave")); # num > 1024
124 &mov ("edi",&wparam(0)); # rp
125 &mov ("eax",&wparam(1)); # ap
126 &mov ("ebx",&wparam(2)); # bp
127 &mov ("edx",&wparam(3)); # np
128 &mov ("esi",&wparam(4)); # n0
129 &mov ("esi",&DWP(0,"esi")); # *n0
131 &lea ("ecx",&DWP($pad,"","ecx",4)); # ecx becomes vector size in bytes
132 &lea ("ebp",&DWP(64,"","ecx",4)); # allocate 4 vectors + 64 bytes
135 &and ("ebp",-64); # align to cache-line
136 &xchg ("ebp","esp"); # alloca
138 &mov ($rp,"edi"); # save rp
139 &mov ($sp,"ebp"); # save esp
141 &mov ($mZeroPrime,"esi");
142 &lea ("esi",&DWP(64,"esp")); # tp
144 &lea ("edi",&DWP(32,"esp")); # scratch area
145 &mov ($scratch,"edi");
148 &lea ("ebp",&DWP(-$pad,"ecx"));
149 &shr ("ebp",2); # restore original num value in ebp
151 &add ("ecx",32/4); # (4 vectors + 32 byte scratch)/4
153 &data_byte(0xf3,0xab); # rep stosl, bzero
156 &lea ("edi",&DWP(64+$pad,"esp","ecx",4));# pointer to ap copy
158 &data_byte(0xf3,0xa5); # rep movsl, memcpy
160 # edi points at the end of ap copy...
162 &add ("edi",$pad); # skip padding to point at bp copy
165 &data_byte(0xf3,0xa5); # rep movsl, memcpy
167 # edi points at the end of bp copy...
169 &add ("edi",$pad); # skip padding to point at np copy
172 &data_byte(0xf3,0xa5); # rep movsl, memcpy
174 # let magic happen...
178 &shl ("ecx",5); # convert word counter to bit counter
180 &data_byte(0xf3,0x0f,0xa6,0xc0);# rep montmul
183 &xor ("edx","edx"); # i=0
184 &lea ("esi",&DWP(64,"esp")); # tp
185 # edi still points at the end of np copy...
187 &lea ("ebp",&DWP(0,"edi","ebp",4)); # so just "rewind"
188 &mov ("edi",$rp); # restore rp
190 &mov ("ebx",&DWP(0,"esi","ecx",4)); # upmost overflow bit
191 &cmp ("ebx",0); # clears CF unconfitionally
192 &jnz (&label("sub"));
193 &mov ("eax",&DWP(-4,"esi","ecx",4));
194 &cmp ("eax",&DWP(-4,"ebp","ecx",4)); # tp[num-1]-np[num-1]?
195 &jae (&label("sub")); # if taken CF is cleared
197 &set_label("copy",4);
199 &data_byte(0xf3,0xa5); # rep movsl
201 &jmp (&label("zap"));
203 &set_label("sub",16);
204 &mov ("eax",&DWP(0,"esi","edx",4));
205 &sbb ("eax",&DWP(0,"ebp","edx",4));
206 &mov (&DWP(0,"edi","edx",4),"eax"); # rp[i]=tp[i]-np[i]
207 &lea ("edx",&DWP(1,"edx")); # i++
208 &dec ("ecx"); # doesn't affect CF!
210 &sbb ("ebx",0); # upmost overflow is still there
212 &jc (&label("copy"));
217 &lea ("ecx",&DWP(64/4+$pad,"","ecx",4));# size of frame divided by 4
219 &data_byte(0xf3,0xab); # rep stosl, bzero
222 &inc ("eax"); # signal "done"
225 &function_end($func);